An investigation on a multi-fear recognition model using facial features and survival horror games

Date of Publication

2015

Document Type

Master's Thesis

Degree Name

Master of Science in Computer Science

Subject Categories

Computer Sciences

College

College of Computer Studies

Department/Unit

Computer Science

Thesis Adviser

Rafael Cabredo

Defense Panel Chair

Jocelynn Cu

Defense Panel Member

Solomon See

Abstract/Summary

This research built fear recognition models in the context of survival horror games. By taking in the visual element of facial features, models able to recognize the fear emotion of the players were achieved. This research aimed to build user-specific models as a first step into a fear recognition model of this context. An affect annotation tool was built by (Vachiratamporn, Legaspi, Moriyama, & Numao,2013) in order to annotate the fear of players based on gameplay during a survival horror game. Data was collected through test subjects video recorded while playing the survival horror game Slender: The Eight Pages (STEP). The spontaneous or induced fear emotions captured from the gameplay were used in order to build the models. Using the annotated fear emotions from the video recording of the player and the facial feature fiducial points, models were built using Support Vector Machines (SVM) in order to recognize the fear emotions (Fear vs. Non-Fear) of a player. The models achieved accuracies of 70-80% with 0.5-0.6 Kappa for 3 Subjects while the lowest of 60% and 0.2 Kappa for the 4th subject. Investigations into modeling the different dimensions of fear (Anxiety, Suspense) and the different intensities (Low,Mid,High) were done, however these yielded poor results of around 30-40% with 0.1-0.2 Kappa. This is due to the overlaps in the data, how similar facial features for Low-Fear also are existent in other fears such as Anxiety and Suspense and the like. The significance of this research is the discovery that having too many emotion labels is detrimental to the study and should be limited to more specific emotions that can be easier annotated by the players. Fear recognition can be achieved using what players annotated as the higher levels of fear, which can be recognized as fear as opposed to non-fear or the other labels such as anxiety and suspense.

Abstract Format

html

Language

English

Format

Electronic

Accession Number

CDTG006579

Shelf Location

Archives, The Learning Commons, 12F Henry Sy Sr. Hall

Physical Description

1 computer optical disc ; 4 3/4 in.

Keywords

Emotion recognition

Upload Full Text

wf_no

This document is currently not available here.

Share

COinS